Root/Documentation/scheduler/sched-design-CFS.txt

1                      =============
2                      CFS Scheduler
3                      =============
4
5
61. OVERVIEW
7
8CFS stands for "Completely Fair Scheduler," and is the new "desktop" process
9scheduler implemented by Ingo Molnar and merged in Linux 2.6.23. It is the
10replacement for the previous vanilla scheduler's SCHED_OTHER interactivity
11code.
12
1380% of CFS's design can be summed up in a single sentence: CFS basically models
14an "ideal, precise multi-tasking CPU" on real hardware.
15
16"Ideal multi-tasking CPU" is a (non-existent :-)) CPU that has 100% physical
17power and which can run each task at precise equal speed, in parallel, each at
181/nr_running speed. For example: if there are 2 tasks running, then it runs
19each at 50% physical power --- i.e., actually in parallel.
20
21On real hardware, we can run only a single task at once, so we have to
22introduce the concept of "virtual runtime." The virtual runtime of a task
23specifies when its next timeslice would start execution on the ideal
24multi-tasking CPU described above. In practice, the virtual runtime of a task
25is its actual runtime normalized to the total number of running tasks.
26
27
28
292. FEW IMPLEMENTATION DETAILS
30
31In CFS the virtual runtime is expressed and tracked via the per-task
32p->se.vruntime (nanosec-unit) value. This way, it's possible to accurately
33timestamp and measure the "expected CPU time" a task should have gotten.
34
35[ small detail: on "ideal" hardware, at any time all tasks would have the same
36  p->se.vruntime value --- i.e., tasks would execute simultaneously and no task
37  would ever get "out of balance" from the "ideal" share of CPU time. ]
38
39CFS's task picking logic is based on this p->se.vruntime value and it is thus
40very simple: it always tries to run the task with the smallest p->se.vruntime
41value (i.e., the task which executed least so far). CFS always tries to split
42up CPU time between runnable tasks as close to "ideal multitasking hardware" as
43possible.
44
45Most of the rest of CFS's design just falls out of this really simple concept,
46with a few add-on embellishments like nice levels, multiprocessing and various
47algorithm variants to recognize sleepers.
48
49
50
513. THE RBTREE
52
53CFS's design is quite radical: it does not use the old data structures for the
54runqueues, but it uses a time-ordered rbtree to build a "timeline" of future
55task execution, and thus has no "array switch" artifacts (by which both the
56previous vanilla scheduler and RSDL/SD are affected).
57
58CFS also maintains the rq->cfs.min_vruntime value, which is a monotonic
59increasing value tracking the smallest vruntime among all tasks in the
60runqueue. The total amount of work done by the system is tracked using
61min_vruntime; that value is used to place newly activated entities on the left
62side of the tree as much as possible.
63
64The total number of running tasks in the runqueue is accounted through the
65rq->cfs.load value, which is the sum of the weights of the tasks queued on the
66runqueue.
67
68CFS maintains a time-ordered rbtree, where all runnable tasks are sorted by the
69p->se.vruntime key (there is a subtraction using rq->cfs.min_vruntime to
70account for possible wraparounds). CFS picks the "leftmost" task from this
71tree and sticks to it.
72As the system progresses forwards, the executed tasks are put into the tree
73more and more to the right --- slowly but surely giving a chance for every task
74to become the "leftmost task" and thus get on the CPU within a deterministic
75amount of time.
76
77Summing up, CFS works like this: it runs a task a bit, and when the task
78schedules (or a scheduler tick happens) the task's CPU usage is "accounted
79for": the (small) time it just spent using the physical CPU is added to
80p->se.vruntime. Once p->se.vruntime gets high enough so that another task
81becomes the "leftmost task" of the time-ordered rbtree it maintains (plus a
82small amount of "granularity" distance relative to the leftmost task so that we
83do not over-schedule tasks and trash the cache), then the new leftmost task is
84picked and the current task is preempted.
85
86
87
884. SOME FEATURES OF CFS
89
90CFS uses nanosecond granularity accounting and does not rely on any jiffies or
91other HZ detail. Thus the CFS scheduler has no notion of "timeslices" in the
92way the previous scheduler had, and has no heuristics whatsoever. There is
93only one central tunable (you have to switch on CONFIG_SCHED_DEBUG):
94
95   /proc/sys/kernel/sched_min_granularity_ns
96
97which can be used to tune the scheduler from "desktop" (i.e., low latencies) to
98"server" (i.e., good batching) workloads. It defaults to a setting suitable
99for desktop workloads. SCHED_BATCH is handled by the CFS scheduler module too.
100
101Due to its design, the CFS scheduler is not prone to any of the "attacks" that
102exist today against the heuristics of the stock scheduler: fiftyp.c, thud.c,
103chew.c, ring-test.c, massive_intr.c all work fine and do not impact
104interactivity and produce the expected behavior.
105
106The CFS scheduler has a much stronger handling of nice levels and SCHED_BATCH
107than the previous vanilla scheduler: both types of workloads are isolated much
108more aggressively.
109
110SMP load-balancing has been reworked/sanitized: the runqueue-walking
111assumptions are gone from the load-balancing code now, and iterators of the
112scheduling modules are used. The balancing code got quite a bit simpler as a
113result.
114
115
116
1175. Scheduling policies
118
119CFS implements three scheduling policies:
120
121  - SCHED_NORMAL (traditionally called SCHED_OTHER): The scheduling
122    policy that is used for regular tasks.
123
124  - SCHED_BATCH: Does not preempt nearly as often as regular tasks
125    would, thereby allowing tasks to run longer and make better use of
126    caches but at the cost of interactivity. This is well suited for
127    batch jobs.
128
129  - SCHED_IDLE: This is even weaker than nice 19, but its not a true
130    idle timer scheduler in order to avoid to get into priority
131    inversion problems which would deadlock the machine.
132
133SCHED_FIFO/_RR are implemented in sched_rt.c and are as specified by
134POSIX.
135
136The command chrt from util-linux-ng 2.13.1.1 can set all of these except
137SCHED_IDLE.
138
139
140
1416. SCHEDULING CLASSES
142
143The new CFS scheduler has been designed in such a way to introduce "Scheduling
144Classes," an extensible hierarchy of scheduler modules. These modules
145encapsulate scheduling policy details and are handled by the scheduler core
146without the core code assuming too much about them.
147
148sched_fair.c implements the CFS scheduler described above.
149
150sched_rt.c implements SCHED_FIFO and SCHED_RR semantics, in a simpler way than
151the previous vanilla scheduler did. It uses 100 runqueues (for all 100 RT
152priority levels, instead of 140 in the previous scheduler) and it needs no
153expired array.
154
155Scheduling classes are implemented through the sched_class structure, which
156contains hooks to functions that must be called whenever an interesting event
157occurs.
158
159This is the (partial) list of the hooks:
160
161 - enqueue_task(...)
162
163   Called when a task enters a runnable state.
164   It puts the scheduling entity (task) into the red-black tree and
165   increments the nr_running variable.
166
167 - dequeue_tree(...)
168
169   When a task is no longer runnable, this function is called to keep the
170   corresponding scheduling entity out of the red-black tree. It decrements
171   the nr_running variable.
172
173 - yield_task(...)
174
175   This function is basically just a dequeue followed by an enqueue, unless the
176   compat_yield sysctl is turned on; in that case, it places the scheduling
177   entity at the right-most end of the red-black tree.
178
179 - check_preempt_curr(...)
180
181   This function checks if a task that entered the runnable state should
182   preempt the currently running task.
183
184 - pick_next_task(...)
185
186   This function chooses the most appropriate task eligible to run next.
187
188 - set_curr_task(...)
189
190   This function is called when a task changes its scheduling class or changes
191   its task group.
192
193 - task_tick(...)
194
195   This function is mostly called from time tick functions; it might lead to
196   process switch. This drives the running preemption.
197
198 - task_new(...)
199
200   The core scheduler gives the scheduling module an opportunity to manage new
201   task startup. The CFS scheduling module uses it for group scheduling, while
202   the scheduling module for a real-time task does not use it.
203
204
205
2067. GROUP SCHEDULER EXTENSIONS TO CFS
207
208Normally, the scheduler operates on individual tasks and strives to provide
209fair CPU time to each task. Sometimes, it may be desirable to group tasks and
210provide fair CPU time to each such task group. For example, it may be
211desirable to first provide fair CPU time to each user on the system and then to
212each task belonging to a user.
213
214CONFIG_GROUP_SCHED strives to achieve exactly that. It lets tasks to be
215grouped and divides CPU time fairly among such groups.
216
217CONFIG_RT_GROUP_SCHED permits to group real-time (i.e., SCHED_FIFO and
218SCHED_RR) tasks.
219
220CONFIG_FAIR_GROUP_SCHED permits to group CFS (i.e., SCHED_NORMAL and
221SCHED_BATCH) tasks.
222
223At present, there are two (mutually exclusive) mechanisms to group tasks for
224CPU bandwidth control purposes:
225
226 - Based on user id (CONFIG_USER_SCHED)
227
228   With this option, tasks are grouped according to their user id.
229
230 - Based on "cgroup" pseudo filesystem (CONFIG_CGROUP_SCHED)
231
232   This options needs CONFIG_CGROUPS to be defined, and lets the administrator
233   create arbitrary groups of tasks, using the "cgroup" pseudo filesystem. See
234   Documentation/cgroups/cgroups.txt for more information about this filesystem.
235
236Only one of these options to group tasks can be chosen and not both.
237
238When CONFIG_USER_SCHED is defined, a directory is created in sysfs for each new
239user and a "cpu_share" file is added in that directory.
240
241    # cd /sys/kernel/uids
242    # cat 512/cpu_share # Display user 512's CPU share
243    1024
244    # echo 2048 > 512/cpu_share # Modify user 512's CPU share
245    # cat 512/cpu_share # Display user 512's CPU share
246    2048
247    #
248
249CPU bandwidth between two users is divided in the ratio of their CPU shares.
250For example: if you would like user "root" to get twice the bandwidth of user
251"guest," then set the cpu_share for both the users such that "root"'s cpu_share
252is twice "guest"'s cpu_share.
253
254When CONFIG_CGROUP_SCHED is defined, a "cpu.shares" file is created for each
255group created using the pseudo filesystem. See example steps below to create
256task groups and modify their CPU share using the "cgroups" pseudo filesystem.
257
258    # mkdir /dev/cpuctl
259    # mount -t cgroup -ocpu none /dev/cpuctl
260    # cd /dev/cpuctl
261
262    # mkdir multimedia # create "multimedia" group of tasks
263    # mkdir browser # create "browser" group of tasks
264
265    # #Configure the multimedia group to receive twice the CPU bandwidth
266    # #that of browser group
267
268    # echo 2048 > multimedia/cpu.shares
269    # echo 1024 > browser/cpu.shares
270
271    # firefox & # Launch firefox and move it to "browser" group
272    # echo <firefox_pid> > browser/tasks
273
274    # #Launch gmplayer (or your favourite movie player)
275    # echo <movie_player_pid> > multimedia/tasks
276
2778. Implementation note: user namespaces
278
279User namespaces are intended to be hierarchical. But they are currently
280only partially implemented. Each of those has ramifications for CFS.
281
282First, since user namespaces are hierarchical, the /sys/kernel/uids
283presentation is inadequate. Eventually we will likely want to use sysfs
284tagging to provide private views of /sys/kernel/uids within each user
285namespace.
286
287Second, the hierarchical nature is intended to support completely
288unprivileged use of user namespaces. So if using user groups, then
289we want the users in a user namespace to be children of the user
290who created it.
291
292That is currently unimplemented. So instead, every user in a new
293user namespace will receive 1024 shares just like any user in the
294initial user namespace. Note that at the moment creation of a new
295user namespace requires each of CAP_SYS_ADMIN, CAP_SETUID, and
296CAP_SETGID.
297

Archive Download this file



interactive